Deep learning-enabled temporal sequencing of metasurface for rewritable and customizable electromagnetic illusions

The realization of rewritable and customizable electromagnetic illusions fundamentally hinges upon the ability to achieve precise spatiotemporal control over electromagnetic waves. Conventional metasurfaces are confined to globally stationary, periodic protocols, lacking the information entropy to orchestrate complex illusion patterns. Here, we introduce a modular metasurface time-domain programming framework that organizes discrete temporal modulation waveforms as reusable modulation units within a predefined library. By flexibly selecting and sequencing these units, rather than redesigning the entire control law for each new task, the framework redistributes the spectral components of the scattered field to synthesize diverse electromagnetic illusions. A deep generative model is established to map target illusions directly to specific time-domain modulation sequences. The metasurface executes these signals across distinct pulses to physically realize the user-defined illusion. Validated on a synthetic aperture imaging testbed, the system rewrites periodic baselines and synthesizes representative aperiodic illusion patterns under user-specified inputs, achieving high fidelity between intended objectives and measurements (structural similarity index ≥0.91). This work establishes a practical route from target-scene specification to executable metasurface control and provides a scalable paradigm for task-driven wave manipulation in radar imaging scenarios.